XIVER, formerly the Philips MEMS foundry · 2025
Text-to-BI for a MEMS foundry
Siloed enterprise data turned into answers people can ask for in plain language, with safety filters and an audit trail. Analytics turnaround cut by 56% and 70% of manual workflows automated.
56%faster analytics turnaround
The problem
XIVER, the MEMS foundry Philips sold to Dutch investors, started with siloed enterprise data, no analytics platform and 23+ stakeholders with different ideas of what AI should do. Any generative AI on that data had to meet ISO and GDPR from day one.
YouWhich process steps drove yield loss last week?
Etch accounts for 42% of last week's yield loss, mostly from two lots on Tuesday.
Etch
Litho
Deposition
Test
Generated query
SELECT step, SUM(scrap_qty) / SUM(start_qty) FROM lot_history JOIN process_steps USING (step_id) WHERE week = 2025-W14 GROUP BY step
My role
Head of Digital Innovation, leading a team of 7
Results
- 70%
- of manual workflows automated
- 23+
- stakeholders across operations, product and growth
Stack
- LangChain
- Azure OpenAI
- Python
- Azure
- SAP S/4HANA
What changes
Getting answers
WithoutA business question waits in the analytics queue.
With Text-to-BI for a MEMS foundryAsked in plain language, answered from company data. Turnaround 56% faster.
Siloed data
WithoutThe same thing has a different name in every system.
With Text-to-BI for a MEMS foundryOne data dictionary and vector retrieval across the sources.
Compliance
WithoutGenerative AI on company data, with no record of what it was asked or told.
With Text-to-BI for a MEMS foundrySafety filters and audit trails, aligned with ISO and GDPR.
Manual work
WithoutTeams rebuild the same reports by hand.
With Text-to-BI for a MEMS foundry70% of manual workflows automated by teams trained to build and evaluate prompts.
How it works
At a high level. A detailed walkthrough is available on a call.
- 01Sources
- 02Dictionary
- 03Retrieve
- 04Generate
- 05Guard
- 06Audit
Decision 01
A shared data dictionary before any model.
A question returns the same answer whichever system the data came from.
Decision 02
Retrieval over company data instead of training a model on it.
Answers stay grounded in current data and point back to their source.
Decision 03
Safety filters and audit trails inside the GenAI framework, not added later.
ISO-aligned data stewardship and GDPR were requirements from the first release.
Decision 04
Teams trained in prompt engineering and model evaluation.
The people closest to the work automated 70% of manual workflows themselves.